Improving our risk communication: Non

Download Report

Transcript Improving our risk communication: Non

Advancing our understanding of
risk assessment: What is it, what
information does it provide, and
how do we talk about its
accuracy?
L. Maaike Helmus
Global Institute of Forensic Research
[email protected]
Cle Ellum, Washington, March 4, 2016
WATSA
Overview
• Nature of risk assessment
• Nature of risk factors
• What information is provided by risk scales?
– Universal risk categories?
– Do people even understand this information?
• How do we assess the accuracy of risk scales?
– Possible statistics (pros/cons)
– My recommendations
• How do we assess change?
• Future directions
What is Risk Assessment?
Diagnosis vs Prognosis
• Diagnosis
– Detecting presence/absence of a condition
• Dichotomous decision
– True state of affairs currently exists
• Prognosis
– Predicting likelihood of an event in the future
– No true state of affairs
– Probabilistic
Diagnosis vs Prognosis
• Different levels of certainty
– Diagnosis: AUC of .80 is ‘good’
– Prognosis: AUCs of .71 and above are large
effect sizes (Rice & Harris, 2005) –
correspond to large Cohen’s d
• Different ways of communicating risk
– 40% chance of being pregnant versus 40%
chance of survival following chemo
Risk Assessment is a task of
Prognosis
• Prediction about future event that may or
may not occur
• Risk is continuous dimension
• Dichotomous predictions have no role in
risk assessment (e.g., ATSA, 2014)
– Cumulative stochastic model
Warnings in Research Studies
• Techniques borrowed from diagnostic field
• Should it apply to risk assessment?
– Similarities: Both examine dichotomous
variable (diseased vs non-diseased; recidivist
versus non-recidivist)
– Differences: Classification versus Prediction.
• Disease is existing state. Recidivism is future
possibility.
Norm-Referenced Scales
–Classical scale construction in
psychology
–Designed to measure varying amounts
of a single construct (e.g., intelligence,
extroversion, psychopathy)
• Factor analyses to better understand construct
–Validity increases with more items
assessing same construct
Criterion-Referenced Scales
– Designed to predict an outcome. Goal:
maximize accuracy
– Often atheoretical
– Measure as many diverse constructs as
possible (incremental validity)
• Items measuring the same construct
considered inefficient
– Antithetical to internal reliability
– Factor structure difficult to detect
What is Risk Assessment?
• Prognostic task
• Measures continuous dimension
• Criterion-referenced scale
Different Approaches to Risk
Assessment
Generations of Risk
Assessment (Bonta, 1996;
Andrews et al., 2006)
• First generation: Unstructured clinical
judgement
• Second generation: Static actuarial
• Third generation: Dynamic actuarial
• Fourth generation: Dynamic actuarial case
management/reassessment
Other Nuances
• Where does SPJ fit in?
– Andrews et al. (2006): variation of first
generation
– Items developed based on research, theory,
OR clinical experience
• What are mechanical vs actuarial scales?
– Hanson & Morton-Bourgon (2009)
– Actuarial: recidivism probability estimates
linked to total scores, items empirically
derived
– Mechanical: no recidivism probabilities, items
may be developed from theory
Types of Risk Assessment
Type of
Evaluation
Unstructured Clinical
Judgement
Empirical-Actuarial
Structured
Professional
Judgement
Mechanical
Factors
?
Overall
Recidivism
Evaluation Estimates
Professional
Judgement
No
Empirically Mechanical
Derived
Actuarial
Yes
No
Theory
Professional
Judgement
Theory
Mechanical
No
SVR-20/HCR-20 (add items)
SRA/STABLE-2000
14
Risk Scales: Accuracy for
Sexual Recidivism
Measures Designed
for Sexual Recidivism
d (95% CI)
N (k)
Empirical Actuarial
.67 (.63-.72)
Mechanical
.66 (.58-.74)
Structured Judgement
.46 (.29-.62)
Unstructured
.42 (.32-.51)
24,089
(81)
5,838
(29)
1,131
(6)
6,456
(11)
15
Hanson & Morton-Bourgon (2009)
Actuarial vs. Professional
Judgment
• Large literature: cognitive psychology,
medicine, weather forecasts, forensic
assessments
• Actuarial outperforms professional
judgement, but only under some
conditions
Good expert
performance
•
•
•
•
•
•
•
•
•
•
Weather forecasters
Livestock & soil judges
Astronomers
Test pilots
Physicists/mathematicians
Chess masters
Accountants
Grain inspectors
Insurance analysts
Photo interpreters
Poor expert
performance
•
•
•
•
•
•
•
•
•
•
•
Clinical psychologists
Psychiatrists
Astrologers
Student admissions
evaluators
Court judges
Behavioural researchers
Counsellors
Personnel selectors
Police officers
Polygraph judges
Stock brokers
When do experts outperform actuarials?
• Shanteau (1992)
– Physical phenomena (not human behaviour)
– Non-unique tasks
– When feedback is available
– Environment tolerant of error
• Kahneman & Klein (2011)
– Environment highly regular (i.e., predictable)
– Expert has considerable practice
– Timely feedback
What types of risk factors might
be included in risk scales?
Static
•
•
•
•
Historical/unchanging
Most reliably documented
Often scored quickly and easily
Tend to be the most predictive
Dynamic
• Stable
– Stable personality characteristics
– Change slowly (if at all)
• Acute
– Rapidly changing
Protective?
• Losel & Farrington (2012)
– Direct protective: factors associated with
lower levels of violence
– Buffering protective: interacts with risk factor
to decrease its influence on recidivism
• Farrington & Ttofi (2011)
– Protective: basically buffering protective
– Promotive: a bit more complicated.
• Each factor has three levels: low, moderate, high
Risk and Promotive Factors
Protective/promotive
• Promotive: Cannot establish without nonarbitrary scaling of low/moderate/high
– Does not exist in our field
• Protective – interaction can apply to risk
factors as well
– Interactions to reduce effect of another risk
factor
Are static/dynamic/protective
different constructs?
• I’m not convinced they are
• Protective can often be reversed and
reframed as a risk factor (Harris & Rice,
2015)
Psychologically Meaningful Risk
Factors
• Risk-relevant propensity (personality
characteristic, latent/underlying trait)
– Mann, Hanson, & Thornton (2010)
• Can be assessed with static, dynamic, or
protective factors
Male Victims
(+)
Lived with a Lover
Deviant Sexual Interests
(-)
Deviant PPG
(+)
Sexual Recidivism
Total Prior Offences
(+)
Negative Attitudes
Towards Supervision
(+)
Antisocial Orientation
Employed
(-)
Value added?
• Yes. Incremental validity of static,
dynamic, and protective factors
• Does not mean they are different
constructs
• More comprehensive assessment of riskrelevant propensities
Other advantages to
distinguishing types of risk factors
• Clinical/conceptual
– What can change, what can’t, positive
psychology (strengths)
• Types of information used
– Static: criminal history info
– Dynamic: interviews, detailed file review,
specialized testing
• More intensive
Brain Break!
What information is provided by
risk assessment scales?
Information Provided by Risk
Scales
• Qualitative
– Nominal risk categories
• Quantitative
– Percentiles
– Risk ratios
– Absolute recidivism estimates
Discrimination vs Calibration (Gail
& Pfeiffer, 2005)
• Discrimination (a.k.a. relative risk)
– Offender’s risk relative to other offenders
– Ranking offenders from highest risk to lowest
risk
– Percentiles, risk ratios
• Calibration (a.k.a. absolute risk)
– Expected probability of recidivism for a
particular score
33
Percentiles
– The commonness or unusualness of a
particular score
• Proportion expected to score higher; lower; the
same
• E.g., “By the end of this presentation, you will learn
that I score in the bottom 5% of researchers in terms
of my ability to make graphs”
– Ideal for norm-referenced constructs
• intelligence, achievement motivation, ego strength,
school grades, athletic competitions
– By far the most commonly used metric in
psychology
• IQ scores, T-scores
Advantages of Percentiles
• Easily understood
• Easily calculated
• Applies to all risk assessment procedures
– Unstructured professional judgement,
structured professional judgement, empirical
actuarial, etc.
• Sufficient for resource allocation decisions
– Priority for treatment
– Surveillance and Monitoring
35
Disadvantages of Percentiles
• Risk tools are criterion referenced, not
norm referenced
• Percentile metric is unlikely to directly
correspond to latent dimension of risk
36
Static-99R scores and relative risk
(log odds ratios)
3.5
3
2.5
2
1.5
1
Static-99R Score
0.5
0
-3 -2 -1 0
-0.5
-1
-1.5
1
2
3
4
5
6
7
8
9 10 11
Based on Helmus et al
K = 21; N = 5,692
37
Static-99R Percentiles and relative
risk (log odds ratios)
3.5
3
2.5
2
1.5
1
Static-99R Score
0.5
0
-0.5
-1
-1.5
0
10 20
30 40
50 60
70 80
90 100
Based on Hanson,
Lloyd, Helmus &
Thornton (k = 4; n =
2,011)
38
Risk Ratios
– How different is this offender from the
typical offender?
– Can be describe using rate ratio, odds
ratio, hazard ratio
• This offender is 2.5 times more likely to
reoffend compared to the typical offender
Advantages of Risk Ratios
• Meaningful reference category
• Inform decisions
– Resource allocation (e.g., treatment or
supervision intensity)
• Risk scales are inherent measures of
relative risk
– Higher scores indicate greater risk, but
how risky?
• More stable than absolute recidivism rates
• More information than percentile rank
40
Disadvantages of Risk Ratios
• Cannot be interpreted in the absence of base
rates
– 2.5 times more likely… than what?
• 10% vs. 50% base rate
• Expected recidivism rates (= risk ratio*Base rate)
only matches absolute recidivism rates in certain
instances
– Low base rate samples
• Hard to understand! (or unwilling?)
– Varela et al. (2014)
41
Absolute Risk
• Expected probability of recidivism
– E.g., This offender scores a 4 on my risk
scale. Other individuals with the same score
have been found to violently reoffend at a rate
of 27% over 5 years
• Unique to actuarial risk scales
Advantages of Absolute Risk
– Available for most of the commonly used
actuarial scales (e.g., MnSOST-R, LSI,
VRAG/SORAG, Static-99R)
– Commonly reported in forensic reports
– Easily understood
– Criterion-referenced
– Important in high-stakes contexts
• Civil commitment in US
• Dangerous offender hearings in Canada?
43
Disadvantages of Absolute Risk
• Hard to specify!
– Significant variability across samples
(Helmus et al., 2012)
– Can change with differences across
time, jurisdiction, policy
– Requires explicit definition
• Outcome
• Length of follow-up
Nominal Risk Categories
• “Low,” “Moderate,” “High”
– Preferred by forensic evaluators and
decision-makers
– Link to action in specific setting
• But what do they mean?
– Inconsistent meanings
• Evaluators use “Low” and “High” risk to
mean different things
• Infer different recidivism probabilities
45
Goal: Develop non-arbitrary
meanings for risk categories
Risk Categories That Work
• Describes people (not risk scales)
– Characteristics of the offender (psychologically
meaningful)
• Tell us what to do
– Linked to realistic options for action
• Evidence-based, scientifically credible
– Universal – applicable to all risk scales
• Simple
– Common Professional Language
• Easy to implement
– Can be done across jurisdictions/scales/offenders
47
Meaningful (perceptible)
differences between categories
• Profile
– Needs: Criminogenic & Non-criminogenic
– Personal and social resources
– Life problems
• Correctional Strategies & Responses
– Human Services
– Supervision
– Custody
• Outcomes
– Base Rate of re-offending
– Expected outcomes with appropriate services
– Life course adjustment
48
Council of State Government Justice Center
Standardized Risk Levels
Level I
Level II
Level III
Level IV
Level V
49
Level I
• Profile:
–
–
–
–
Criminogenic needs: none or transitory
Non-Criminogenic needs: none or transitory
Resources: clearly identifiable personal and social resources
Reoffending Base Rate: same as non-offender community at
large (e.g., <5%)
• Sex offenders: similar to non-sex offenders committing a sex offence (<
2%)
• Correctional Strategies:
– Human services: unnecessary/ direct to social services if desired
– Community Supervision: minimal as not necessary for compliance
– Custody: counterproductive
• Outcomes:
– Short-term: No change in probability of re-offending
– Life Course: Desistance
50
Level II
• Profile:
– Criminogenic needs: Few, if any, more acute than chronic.
– Non-Criminogenic needs: Few if any, transitory and quick to
respond
– Resources: Clearly identifiable personal and social resources
– Reoffending Base Rate: Higher than individuals without a
criminal record but lower than typical offender
• Correctional Strategies:
– Human services: Brief interventions: social problem solving
aided through existing community services.
– Community Supervision: simple case management with minimal
supervision for compliance
– Custody: may be productive for crisis management but
unnecessary long-term
• Outcomes:
– Short-term: Probability of re-offending reduces quickly to nonoffender levels (Level I)
51
– Life Course: Desistance (i.e., one and done)
Level III
• Profile:
– Criminogenic needs: Some; may be severe in one or two discrete needs or
less chronic/severe across multiple needs
– Non-Criminogenic needs: Some; typical to offenders
– Resources: Some identifiable personal and social resources, sporadic use
– Reoffending Base Rate: Same as the average rate for offenders as a group
• Correctional Strategies:
– Human services: Structured services target criminogenic needs over
months; (e.g. ~ 100-200 hours of service); assist with non-criminogenic
needs/responsivity factors.
– Community Supervision: Change focused supervision with supervision
for enhance engagement and compliance
– Custody: may support short-term risk management
• Outcomes
– Short-term: Probability of re-offending can significantly ↓ with
appropriate strategies BUT still higher than community at large (Level II)
– Life Course: Expected involvement in crime over time but desistance
possible over years
52
Level IV
• Profile:
– Criminogenic needs: Multiple; may be chronic in one or two discrete needs and
moderate in others
– Non-Criminogenic needs: Multiple, moderate to severe.
– Resources: Few identifiable personal and social resources, sporadic prosocial
use
– Reoffending Base Rate: Higher than the rate of “typical” offender
• Correctional Strategies:
– Human services: Structured comprehensive services target multiple
criminogenic needs over lengthy period with community follow-ups and supports
(e.g. ~ 300+ hours of service); assist with non-criminogenic needs/responsivity
factors.
– Community Supervision: Intensive and change focused supervision with
supervision for enhance engagement and compliance as well as risk
management
– Custody: may be productive for short-term risk management and beginning
treatment
• Outcomes:
– Short-term: Probability of re-offending can ↓ with appropriate strategies BUT still
“average” for offender population at large.
– Life Course: Chronic offending rate shows gradual decline with appropriate
services and time (i.e., years/decades).
53
Level V
•
Profile:
– Criminogenic needs: Multiple, chronic and entrenched
– Non-Criminogenic needs: Multiple, moderate to severe
– Resources: Few identifiable personal and social resources; use for procriminal
pursuits
– Reoffending Base Rate: 85% re-offending rate (Top 5% of offenders)
• Not currently possible to empirically identify this group with sex offenders
•
Correctional Strategies:
– Human services: Structured comprehensive services target multiple criminogenic
needs over years ideally prior to community release (e.g. ~ 300+ hours of service);
assist with non-criminogenic needs/responsivity factors
– Community Supervision: Long-term supervision begins with intensive/risk
management focus and gradual change to change focus (Level IV) with success over
time
– Custody: incapacitation for community safety , long-term change strategy starts with
behavioral management
•
Outcomes
– Short-term: Probability of re-offending still significantly higher than offender
population even with appropriate long-term correctional strategies
– Life Course: Chronic offending rate gradually declines over decades or advanced
age, cascade within custody environments
54
Three Quantitative Indicators
• Absolute recidivism rates
– 5%, 85% reconvicted after 2 years
• Percentile rank
– median – middle risk level (50% score same
or lower)
• Risk Ratios
– 1.4 times as likely to reoffend as those in the
middle of the risk distribution
Calculating Risk Categories 1
Requirement: Risk Scores & Recidivism of Population
Recidivism
1.00
Risk Scores Distribution
Base Rate
~ .40
Median
0.00
Calculating Risk Categories 2
Upper Boundary
~5% Recidivism
Category I: Upper Boundary
Category V: Lower Boundary
Lower Boundary
~85% Recidivism
1.00
~0.85
Cat I
Cat V
~0.40
~0.05
0.00
Category I:
Risk Score
Cutoff
Category V:
Risk Score
Cutoff
Calculating Risk Categories 3
Category III: Boundaries
Boundaries = Appropriate Treatment Effect Size
Odds Ratio: ±1.44
1.00
~0.85
Cat III
Cat II
Cat IV
Cat I
~0.40
Cat V
~0.05
0.00
Category III:
Risk Score
Cutoffs
New STATIC risk categories
• Currently, Static-99R has 4 categories:
– Low, Low-Moderate, Moderate-High, High
• Static-2002R has 5:
– Low, Low-Moderate, Moderate, ModerateHigh, High
• Standardize STATIC categories
Static-99R
Name
Midpoint
Scores percentile
Static-2002R
Predicted
5-year
Recidivism
rate (%)
Scores
Midpoint
percentile
Predicted
5-year
Recidivism
rate (%)
I
Very
Low Risk
-3, -2
2.8
0.9 – 1.3
-2, -1
2.8
1.0 – 1.5
II
Below
Average
-1, 0
14.8
1.9 – 2.8
0,1
13.9
2.2 – 3.2
1, 2, 3
49.1
3.9 – 7.9
2, 3, 4
46.7
4.6 – 9.7
4, 5
83.4
11.0 – 15.2
5, 6
81.6
13.8 – 19.2
6 or
more
96.2
20.5 – 53.0
7 or
more
96.0
26.0 – 53.5
III
IV-a
IV-b
Average
Risk
Above
Average
Well
Above
Average
Comparison of Original and
Revised STATIC categories
Original Category Agreement: 51% (squares)
Revised Category Agreement: 72% (shaded area)
Summary: Information Provided
by Risk Scales
•
•
•
•
•
Total score (actuarial)
Percentile
Risk Ratio
Recidivism Estimate (actuarial)
Nominal Risk Category
• Fuller picture of risk: Use multiple pieces
of information
Do People Understand the Info
Provided by Risk Scales?
Varela, J. G., Boccaccini, M. T., Cuervo,
V. A., Murrie, D. C., & Clark, J. W. (2014).
Same score, different message:
Perceptions of offender risk depend on
Static-99R risk communication format.
Law and Human Behavior, 38, 418-427.
doi:10.1037/lhb0000073
Method
• 211 adult community members called for
jury duty
• 2-page document about case and Static99R
• Manipulations:
– Low score (1) versus high score (6)
– Risk communication format
• risk category (low vs. high)
• risk ratio (three-fourths vs. 2.9 times the recidivism
rate of typical offender)
• absolute recidivism estimate (9.4% or 31.2%)
Outcome Measure
• Participants rated on scale of 1 to 6.
• Low scores = lower perceived risk.
– Likelihood of committing a new sex offence
– Dangerousness to community members
Findings
• When asked to make dichotomous
decision, 95% of participants indicated that
the offender would commit a new sex
offence in the next 5 years
Findings
• Whether participants rated the low risk
offender as lower risk than the high risk
offender depended on how the information
was communicated
– Nominal risk category: low risk offender given
lower scores than high risk offender
– Risk ratio: low risk offenders given nonsignificantly lower scores than high risk
offender (p = .12)
– Absolute recidivism rate: low risk offenders
given same score as high risk
Effect of Communication Metric
• Score of 6
– Those who got the nominal risk category
assigned a higher risk score than those who
got a risk ratio or recidivism estimate
• Score of 1
– Those who got the nominal risk category
assigned the lowest scores, but not
significantly lower than the other formats
Understanding of Risk Ratios
• Message: His risk is 2.9 times higher than
recidivism rate of the typical sex offender
– 85% said he was more likely to reoffend than
the typical sex offender
• Message: His risk is three-quarters the
recidivism rate of the typical sex offender
– 80% said he was more likely to reoffend than
the typical sex offender
How Important/Difficult Were
Static-99R Results?
• They were rated as more important for
higher risk offenders
– When the information was provided as
nominal risk category or risk ratio
• Those who read about low scoring
offenders reported Static-99R as more
difficult to understand
User surveys: What are people
using/liking/understanding?
Blais & Forth (2014)
• 111 reports for DO hearings (court or
prosecution-appointed)
• Over 90% used actuarial scale; 53% SPJ
• PCL-R used in over 95% of reports
• Static-99 was most common risk scale
(60%)
Singh et al. (2014)
• 2,135 mental health professionals doing
violence risk assessment
• Half from Europe, 21% from North
America
• Over 400 instruments used; roughly half
developed for particular institution/setting
• Most common: HCR-20, then PCL-R
– Roughly half were SPJ, half actuarial
Neal & Grisso (2014)
• 434 psychiatrists/psychologists (868
cases)
• Most from US, Canada, Europe, Australia,
New Zealand
• Most common referrals: competence to
stand trial, violence risk, sex offender risk,
insanity, sentencing, disability, child
custody, civil commitment, child protection,
civil tort
Neal & Grisso (2014)
• Structured risk tools
– Least likely for: competence (58%), disability
(66%), civil tort (67%)
– Most likely for sex offender risk (97%), child
protection (93%), and violence risk (89%)
• Sex offender risk: Static-99R/2002R most
common (66%), followed by PCL-R and
personality assessments
Archer et al. (2006)
• Survey of American psychologists
• Sex offenders: Static-99 most common,
followed closely by other scales (SVR-20)
• Similar to findings by Jackson & Hess
(2007; civil commitment) and McGrath et
al (2010; treatment)
Blais & Forth (2014)
• 95% use nominal risk categories
• Actuarial scales
– ~66% report total score
– 90% report recidivism estimate
– 37% report percentile
• SPJ
– 24% report a total score
Chevalier et al. (2014)
• 109 experts using Static-99R in SVP
evaluations
• What do they report?
– 83% report nominal risk categories
– 83% report recidivism estimates
– 35% report percentiles
– 33% report risk ratios
• What’s most important information?
– 54% say recidivism estimates
– 25% say nominal risk categories
How do we assess the
accuracy of risk scales?
Possible Statistics That Could
Be Used
• Singh, J. P. (2013). Predictive validity
performance indicators in violence risk
assessment: A methodological primer.
Behavioral Sciences and the Law, 31, 822
Possible Statistics (Singh, 2013)
•
•
•
•
•
•
•
•
•
•
Sensitivity
Specificity
Positive predictive value
Negative predictive value
Number needed to detain
Number safely discharged
Diagnostic odds ratio
Logistic odds ratio
Point-biserial correlation
AUC (Area under the curve)
Definitions
Reality
Cancerous
(Recidivism)
Diagnosis
Correct!
True positive
Cancerous
(Recidivism)
(hit)
Error!
Not cancerous
False negative
(No recid)
(miss)
Not cancerous
(No recid)
Error!
False positive
(false alarm)
Correct!
True negative
84
Possible Statistics
• Sensitivity
– TP/(TP+FN)
– Proportion of recidivists correctly ‘diagnosed’
as recidivists
• Specificity
– TN/(TN+FP)
– Proportion of non-recidivists correctly
‘diagnosed’ as non-recidivists
Possible Statistics
• Positive Predictive Value
– TP/(TP+FP)
– Proportion of diagnosed recidivists that
actually did recidivate
• Negative Predictive Value
– TN/(TN+FN)
– Proportion of diagnosed non-recidivists that
actually did not recidivate
Possible Statistics
• Number needed to detain
– 1/PPV
– Number of diagnosed recidivists we need to
detain to prevent 1 act of recidivism
• Number safely discharged
– (1/(1-NPV))-1
– Number of diagnosed non-recidivists we could
discharge before a recidivism incident occurs
Possible Statistics
• Diagnostic Odds ratio
– Singh (2013): odds of a TP relative to the
odds of a FP
Problems with These Statistics
• Developed for diagnostic decisions
(dichotomous)
– Not appropriate for prognostic decisions
– Inappropriate to use any risk scale to
classify offenders as recidivists or nonrecidivists
• ‘High risk’ is not synonymous with ‘Going to
recidivate’ (may be less than 50%)
– May just mean: this guy gets more intensive
supervision
Problems with These Statistics
• Base rates below 50%
– If goal is to maximize TP and TN, you should
predict “no” for all offenders
– But what if low risk group has 4% recidivism
rate and high risk group has 40% recidivism
rate?
• Meaningful info for risk management
Problems with These Statistics
• PPV rates determined by overall base rate
– Low base rate: Even with high AUC (>.90),
PPV rate will be low
– Artificial ways to boost your PPV: choose
more common outcome (violence instead of
sex offence), increase follow-up
• Makes scale look more accurate
• Is it?
Possible Statistics
• Log odds ratio
– Expresses how log odds of recidivism
increases with each one-point increase on the
risk scale
• Point-biserial correlation & AUC
– Expresses how recidivism increases with
higher risk scores
Log odds ratios
• Log odds ratios
– Can’t compare for scales with different range
of scores
– Increase in odds of recidivism for each onepoint increase in scale
• Meaning of one point difference varies across
scales (e.g., Static-99R vs PCL-R)
AUCs
• Probability that a randomly selected
recidivist has a higher risk score than a
randomly selected non-recidivist
• Varies between 0 and 1.
– Below .5 is negative accuracy (low risk more
likely to reoffend)
– Above .5 is positive accuracy (high risk more
likely to reoffend)
• Values of .56, .64, and .71 are
low/moderate/high accuracy, respectively
Point-biserial correlations
• Ranges between -1 and +1
• Strongly influenced by recidivism rate
• As base rate deviates from 50%, r gets
smaller
• If recidivism rate is 5% and the scale has
perfect predictive accuracy, r will still be
.47 (Singh, 2013)
– Traditionally, values of .10, .30, and .50 are
considered small/moderate/large
How much is the base rate
going to impact my effect
size?
Babchishin, K. M., & Helmus, L. M. (2015,
Online First). The influence of base rates on
correlations: An evaluation of proposed
alternative effect sizes using real-world
dichotomous data. Behavior Research
Methods.
Possible Statistics (Singh, 2013)
•
•
•
•
•
•
•
•
•
•
Sensitivity
Specificity
Positive predictive value
Negative predictive value
Number needed to detain
Number safely discharged
Diagnostic odds ratio
Logistic odds ratio
Point-biserial correlation
AUC (Area under the curve)
Likelihood ratios
• Mossman (2006); Wollert et al. (2010)
• Unique LR for each score on risk scale
• Probability of having a particular risk score
among recidivists divided by the
probability of having that score among
non-recidivists
Problems with Likelihood Ratios
• Harris & Rice, 2007; Helmus & Thornton,
2014
• Intended for diagnosis tests, not prognosis
• Meant to be invariant to base rate
– Medical context: people who change from
non-diseased to diseased change their
symptoms
– Offenders change from non-recidivist to
recidivist without changing their initial risk
score
• Not stable across follow-up time, even in same
study
Example: Likelihood ratios for each risk/age group on the
MATS-1 scale at each year of follow-up (up to 15 years).
Source: Helmus & Thornton, 2014
Problems with Likelihood ratios
• Artificially affected by distribution of risk
scores
Problems with Likelihood Ratios
Risk
Category
N
Recidivism
2,500
5,000
2,500
10,000
5%
10%
15%
10%
125
500
375
1,000
2,375
4,500
2,125
9,000
1,000
4,000
5,000
10,000
5%
10%
15%
12%
50
400
750
1,200
950
3,600
4,250
8,800
N
N nonrecidivists recidivists
Sample 1
Low
Moderate
High
Total
Sample 2
Low
Moderate
High
Total
Problems with Likelihood Ratios
Risk
Category
N
Recidivism
2,500
5,000
2,500
10,000
5%
10%
15%
10%
125
500
375
1,000
2,375
4,500
2,125
9,000
.47
1.00
1.59
1,000
4,000
5,000
10,000
5%
10%
15%
12%
50
400
750
1,200
950
3,600
4,250
8,800
.39
.81
1.29
N
N nonLikelihood
recidivists recidivists
Ratios
Sample 1
Low
Moderate
High
Total
Sample 2
Low
Moderate
High
Total
Brain Break!
What Statistics Should we
Report?
• My recommendations
Relative Predictive Accuracy
• To assess scale’s ability to discriminate
between recidivists and non-recidivists
– AUCs
• Also: Harrell’s c
– Risk ratios
• Odds ratios from logistic regression
• Hazard ratios from Cox regression
Evaluating Absolute Predictive
Accuracy
• Calibration
• Ignored in offender recidivism prediction
but well developed in other fields (e.g.,
medicine)
• To what extent do the observed values (O)
correspond to the predicted values (E)?
109
Measure of Effect Size
• ER/OR index
– (Number Recidivists Expected)/(Number
Recidivists Observed)
• Poisson variance for the logarithm of the
observed number of cases (OR):
95%CI  ( ER / OR ) * e


 1.96 1 


O
R


110
Interpreting ER/OR
• ER/OR = 1
– Number of recidivists matches predicted number
• ER/OR < 1
– Scale underpredicted recidivism
• ER/OR > 1
– Scale overpredicted recidivism
• 95% CI does not include 1: significant difference
between observed and expected recidivists
111
Recidivism Rates (5 years sex)
Helmus, Thornton et al. (2012)
18
16
14
12
Static-99
Static-99R
Observed
10
8
6
4
2
0
20s
30s
40s
50s
60s
70+
112
ER/OR index – 5 year sex recidivism
Age group
Static-99
Static-99R
20s
0.91
1.03
30s
0.88
1.01
40s
1.16
0.93
50s
1.13
0.91
60s
3.06**
1.49
70s
2.41*
1.20
113
Discussing Accuracy of Risk
Scales
• Consider both relative and absolute accuracy
• Statistics from other fields are useful (e.g.,
medicine)
– Ensure your application matches context in which it
was developed (e.g., diagnosis vs. prognosis)
• I like:
– AUCs and/or risk ratios (odds ratio, hazard ratio)
– E/O index
114
How Can We Assess Offender
Change?
Assessing John
• When John started his community
supervision, his STABLE-2007 score was 9
• One year later, I have re-assessed the
STABLE and he scores 4
• Has John changed?
Basic Data
10
Recidivism
9
8
7
6
Score 5
4
3
2
1
0
0
2
4
6
8
Length of Follow-Up (months)
10
12
Assessing Offender Change
• Classical psychological assessment
– Difference between multiple assessments is
measurement error, not change
• Need to demonstrate reliable change
beyond measurement error
Question for Analyses of Change
• How should we estimate a value (e.g., for
cooperation) at the time of recidivism?
• Imputation essential because we will never
have precise measurements before all
recidivism events.
• Time-invariant survival analysis (the
standard) assumes that initial values
remain unchanged throughout the followup period.
Imputation Option 1:
Last is Best; Use Most Recent
Recidivism
10
9
8
7
6
5
4
3
2
1
0
Cooperation
Observation
0
2
4
6
8
10
12
Imputation Option 2:
Linear: Impossible Results
Observation
Recidivism
10
9
8
7
6
5
4
3
2
1
0
Observation
0
2
4
6
8
10
12
Option 3:
Average of Prior Assessments
Recidivism
10
9
8
7
6
5
4
3
2
1
0
Average
Observation
0
2
4
6
8
10
12
Multiple Assessments
Imputation: Most Recent
Imputation: Moving Average (last 2)
Moving Average (last 3)
Moving Average (cumulative)
Arguments for Some Form of
Average
• Reliability < 1.0
• Regression to the mean
– low scores go up; high scores come down
Future Directions in Assessing
Offender Change
• What predicts best?
– First assessment?
– Last assessment?
– Average assessment?
– Weighted average?
– Linear prediction?
• Statistical analysis techniques
– HLM
– AIC/BIC for comparing non-nested models
Preview of Coming Attractions
Special Issue!
• Criminal Justice and Behavior
– Statistical Issues and Innovations in
Predicting Recidivism
• Edited by me and Kelly Babchishin
Absolute recidivism estimates
• Hard to generate!
• Shakiest of the risk communication metrics
– But among most commonly reported
• Greater research on stability across
samples, identifying and incorporating
sources of variability
Scale Quality: A Tale of Two
States
• Two field studies:
• Texas
– Static-99 AUC = .57 (Boccaccini et al. 2009)
– No information on training, experience, quality
control
• California
– Static-99R AUC > .80
– Most rigorous training and certification we’ve
seen
Quality of Implementation
Matters!
• We know on average, these risk factors
predict
• Need to demonstrate that you’re doing a
reliable, high-quality job of assessing them
• Indicators of risk assessment training and
quality should be more routinely reported
Upcoming Field Studies of
Static-99R
• Will be presented at ATSA 2016
– Texas –approx. 34,000 offenders
– British Columbia –approx. 4,000 offenders
– California – approx. 1,500 offenders
Risk Communication
• Is target audience understanding risk
information?
• Use of graphs, common language,
reframing
• Numeracy
• See 2015 special issue of Behavioral
Sciences and the Law
A glimpse in the future?
Comprehensive risk assessment model
Individual Factors
Initial
Assessment
Environment
Factors
Offender
Change
Revised
Assessment
Environment
Change
Thanks for your time!
Contact: [email protected]